Deslop

Deslop

Slop is the absence of a decision — the statistically likeliest filler standing where a choice should have been made. That is the test for every finding: was this chosen? The same construction, picked deliberately in the author's voice, is not slop. The goal is never "clean" text; it is text where every word survived a decision.

Two facts shape the method:

1. You cannot see your own slop. Slop patterns are artifacts of preference tuning, not vocabulary mistakes — the model that produces them rates them as good writing on re-read. Verification is therefore mechanical (re-scan with the checks in references/checks.md), never impressionistic.
2. Overcorrection is also slop. Scrubbed-flat prose, fake typos, forced slang, and mechanical long/short sentence alternation are their own recognizable genre. Voice is protected before anything is subtracted.

Scope

This skill governs delivered prose — text a reader consumes as writing: copy, posts, essays, emails, marketing, narrative documentation, anything whose job is to be read.

It does not govern working documents — reference files, specs, agent and skill instructions, changelogs, schemas, routing tables, navigation furniture. There a character is doing structural work, not rhetorical work: a · separating two links in a table is not an aside, and a dash setting off a column label is not a dramatic pause. Applying the typographic rules there produces churn without a reader who benefits.

The test: is this read as writing, or read to find something? Only the first is in scope. Everything below — the dash budget and middle-dot ban especially, since they are about a reader's ear for machine prose — assumes the first. This skill's own reference files are out of scope by that rule, and are written accordingly.

Modes

Principles

1. Protect voice first. Before editing, list 3–5 voice signals to preserve (vocabulary level, cadence, bluntness, humor, uncertainty style, digressions, punctuation habits). A user-supplied writing sample overrides every default in this skill, including the dash budget. See references/voice.md.
2. Delete first, rewrite second. Most slop guards nothing — cut it. Rewrite only what carries meaning. Cap total deletion at ~25% of the piece; past that you are gutting content, not removing slop.
3. Rewrite by meaning, never by paraphrasing the pattern. Ask what the sentence is for, then say that plainly. A paraphrase that preserves the move ("the real X is Y" for "not X but Y") is a new finding, not a fix — the banned escape hatches are listed in references/structures.md.
4. Fake-profound lines are deleted, not re-metaphored. Replacing one purple kicker with a better metaphor reproduces the disease.
5. Never invent facts. Specificity comes from the source or from the author. Where a concrete detail is missing, leave [ADD: which study? what year?] — never a plausible-sounding fill. Before delivering, answer explicitly: does the rewrite state any fact, name, number, date, or citation not present in the source?
6. One strong finding beats three weak ones. Bold + scare quotes + a dash aside on the same phrase is one stacked tell, not three. Consolidate before reporting.
7. Every rule carries its exemption. House-style Title Case, genuine tricolons, term-definition bullets in docs, a single "however" — the exemption columns in the reference files are as binding as the rules.
8. Match the register. Reddit, academic, docs, and marketing have different fatal tells and different legitimate structures — check the register guide in references/voice.md before flagging.

Process (Edit mode)

1. Read whole, capture voice. Read the full piece once. Note register + audience. Write down the 3–5 protected voice signals.
2. Scan. Sweep against references/patterns.md (surface tells), references/structures.md (structural moves), and the mechanical checks in references/checks.md. Collect findings with quotes.
3. Triage. Apply exemptions; consolidate stacked tells; drop anything that is a defensible authorial choice or a protected voice signal.
4. Rewrite. Delete-first. Per finding: what is this sentence for? Say that, in the author's voice. No escape-hatch paraphrases.
5. Verify mechanically. Re-run the scan against your own output. Any hit is a new finding. Loop at most 3 passes; if a passage still fires, rewrite it from the bare claim in plain declarative sentences.
6. Fact check. Answer the invented-facts question from Principle 5.
7. Deliver. Final text plus a change note of at most 8 lines, grouped by pattern. If the note needs more than 8 lines, you changed too much — reconsider.

Quick checks (30 seconds, before delivering)

• Any middle-dot separator (·) or em-dash (—)? In delivered prose, never use either; both are top AI tells. Use normal punctuation or layout for lists and navigation. When a dash is genuinely warranted, make it an en-dash (–). Dash budget stays near 0 (≤2 in a long piece); never swap a dash for parentheses, restructure the sentence instead. (Voice sample overrides. Working documents are out of scope — see Scope.)
• Any "not X but Y" — or any of its escape-hatch paraphrases?
• Any inanimate noun performing a human action ("the data tells us")?
• Three items where two would do? Three sentences of the same length in a row?
• Any sentence that sounds like a pull-quote? Delete it.
• Opener that clears its throat ("Here's the thing")? Ending that recaps or reassures?
• Any bold mid-prose, emoji in headings, **Label:** restatement bullets?
• Any fact, number, or name the source doesn't contain?
• Would the author actually say this sentence out loud?

Code mode

Scope: the diff (git diff main or this session's changes), not the whole repo. Keep behavior unchanged; minimal focused edits; match the file's existing style. Remove:

• Comments a human wouldn't write — narrating the obvious, explaining the change to a reviewer ("// added to handle the new case"), restating the next line. Keep only comments stating what the code cannot: constraints, whys, workarounds.
• Defensive bloat — try/catch, null checks, and validation abnormal for that area, especially on trusted/already-validated codepaths; generic except Exception / empty catch blocks.
• Type evasion — any casts and assertions used to silence the checker instead of fixing the type.
• Placeholder naming — data, result, temp, handleData, Manager/Helper suffixes carrying no specificity.
• Ceremony — zero-information docstrings, single-use wrapper layers, speculative abstraction, dead code, debug leftovers, README boilerplate ("This project aims to..."), emoji in code or docs.
• Any other style inconsistent with the file.

Run the tests after. Report in 1–3 sentences.

Credits

Merged and reworked from: blader/humanizer (Siqi Chen; the Wikipedia "Signs of AI writing" lineage — also the root of the Zed, Hermes, jpeggdev, softaworks, and humanizer-zh forks), hardikpandya/stop-slop, juliusbrussee fuck-slop, petergyang/no-ai-slop, jpeggdev/humanize-writing, realrossmanngroup/no_ai_slop_writing_rules, nousresearch/hermes-agent humanizer, worldwonderer/oh-story-claudecode story-deslop, cursor/plugins deslop, and brianlovin/agent-config deslop. All doctrine belongs to its authors; this is a transformative merge with attribution.